Plans for data centers in orbit have shifted from speculation to regulatory filings at startling speed. SpaceX told the FCC in January 2026 it wants to launch up to one million satellites dedicated to AI processing. Blue Origin followed with an application for 51,600 similar units. Google has held talks with SpaceX on the concept while advancing its own Project Suncatcher prototypes slated for 2027 launch. The idea sells itself on paper. Constant sunlight for power. No local zoning fights. Passive cooling into the void of space.
But one obstacle stands out. The very networking problem engineers spent decades solving on the ground now returns in vacuum. TechRadar reported this week that orbital proposals reintroduce latency and predictability issues largely tamed by dense fiber networks and internet exchange points on Earth.
Low-Earth orbit satellites sit only milliseconds away in pure distance. Yet real-world data movement to ground stations or between satellites faces atmospheric interference, handoffs between moving platforms, and the sheer physics of radio or laser links across hundreds of kilometers. Those variables matter when AI inference demands consistent response times. A few dozen milliseconds can separate useful computation from wasted cycles.
DE-CIX CEO Ivo Ivanov put it plainly. “Orbital compute should be viewed as another layer in our increasingly distributed digital ecosystem – not an outright replacement for terrestrial compute.” He added that the real test lies in “creating an interconnection layer that makes orbital, terrestrial, cloud, and edge infrastructure behave as though they’re part of the same ecosystem.” Ivanov’s comments, highlighted in the TechRadar piece, reflect a growing industry view. Space facilities won’t displace ground campuses. They will handle specific workloads where energy abundance outweighs proximity costs.
Terrestrial data centers have hit limits. Power grids strain under AI training clusters that consume dozens of megawatts each. Cooling requires vast water volumes. Communities push back against noise, heat, and transmission lines. A Forbes analysis of a JLL report published June 15 noted backlash against new ground facilities and questions whether terrestrial buildout can match demand. Orbital alternatives promise different math. Solar arrays drink unlimited photons. Heat radiates efficiently into space without fans or chillers. One JLL projection suggested a 40-megawatt orbital setup could save $138 million in energy costs over a decade compared with equivalent Earth operations.
Yet those gains come with trade-offs. Cooling in vacuum relies on large radiators since convection doesn’t exist without air. The Forbes piece explained that such surfaces must grow substantially larger, driving up mass and therefore launch expense. Radiation hardening adds another layer of cost and complexity for processors. Elon Musk addressed some of these points in March. He described SpaceX’s AI Sat Mini design with 100 kilowatts of power, solar arrays stretching longer than 170 meters, and a 100-square-meter radiator. The company aims to scale to megawatt-class satellites. Musk predicted space-based AI would undercut terrestrial costs within two or three years thanks to free solar energy and no real-estate constraints.
Networking remains the uncertain variable. On land, operators perfected dense interconnections. Massive fiber routes link hyperscale campuses to cloud providers with sub-millisecond latency inside metro areas. Orbital constellations must stitch themselves into that fabric using laser inter-satellite links or optical feeder connections to ground. The European Space Agency’s OFELIAS project and DE-CIX collaboration with the German Aerospace Center target stable optical links precisely to reduce variability. Success here determines whether orbital capacity functions as true extension or isolated islands of compute.
Recent conversations show momentum building. TechCrunch reported in May that Google and SpaceX have discussed orbital data centers as SpaceX prepares for a potential $1.75 trillion IPO. The talks follow Google’s earlier $900 million investment in SpaceX and align with broader efforts like Anthropic’s computing partnership with the company. Startups such as Starcloud have trained AI models in orbit already and filed for their own large constellations. Axiom Space deployed early orbital computing nodes in January.
Analysts caution that economics still favor Earth for most tasks. The JLL report cited in Forbes sets a launch cost threshold around $500 per kilogram for competitiveness. Current costs hover near $2,700, though SpaceX’s Starship development targets $200. Until that gap closes, orbital facilities will likely serve niche roles. Edge processing of Earth-observation data offers one immediate application. AI models running close to sensors avoid hauling raw information to ground stations that only see satellites for brief windows each day.
Power presents another contrast. Terrestrial facilities compete for grid capacity already stretched by electrification and renewables integration. Orbital setups generate their own electricity around the clock in sunlight, though they must store or manage it during eclipse periods. Musk has spoken of producing processors at terawatt scale annually through a Terafab project involving SpaceX, Tesla, and xAI. That volume, roughly 50 times current advanced chip output, would supply the processors needed for massive constellations.
Still, questions linger on maintenance, upgrades, and end-of-life. Satellites cannot receive physical repairs. Software updates help, but hardware obsolescence arrives quickly in AI. Constellations must launch replacement waves regularly. Debris management grows critical as thousands or millions of units operate in crowded low-Earth orbits. The Forbes article flagged operational reliability and debris as key risks alongside launch economics.
Industry voices increasingly frame the choice as complementary rather than competitive. Ivanov noted that some AI training might migrate to orbit for its energy profile while latency-sensitive inference stays closer to users on the ground. The future infrastructure, he suggested, will let workloads flow to their most efficient home with data moving predictably across layers. Achieving that vision demands coordinated work between satellite operators, network providers, and standards bodies now.
Recent X discussions echo this mix of excitement and skepticism. One user called orbital data centers “inevitable.” Another argued ocean-sunk facilities make more sense for cooling and cost. Starcloud’s CEO posted images of new ground facilities, drawing comments that such sites will build the thousands of satellites required. Grok itself observed that space offers constant solar power and radiative cooling without water or grids, turning phased constellation launches into evolving infrastructure.
Technical progress continues on multiple fronts. NVIDIA has released space-rated platforms. Radiation-hardened chips are advancing. Laser communication networks proven in Starlink provide a foundation Musk says needs no “magic” for data-center adaptation. Yet the integration challenge TechRadar highlighted persists. Without reliable, low-variability networking, orbital compute risks becoming stranded capacity.
The coming years will test these assumptions. First demonstration satellites could fly as soon as late 2026 or 2027. Regulatory approvals, international spectrum coordination, and actual cost curves will decide pace. For now the bet from SpaceX, Google, and others rests on AI demand outstripping Earth’s practical limits fast enough to justify the orbital detour. Networking, more than rockets or chips, may decide if that bet pays off.
WebProNews is an iEntry Publication